Abstract
Learning studies are often characterized by a lack of generalization to new tasks and stimuli that fundamentally limits the value of training. A long-standing debate is whether the ‘curse’ of specificity is inherent to learning or may result from a lack of diversity in task and stimuli. By varying the diversity of a set of training stimuli, we show that it is possible to induce two different types of learning: (i) specific learning with limited generalization using a restricted training set; (ii) model-based generalizable learning using a diverse training set. These results suggest that the ‘curse’ of specificity can be the fault of the paradigm.
Observers were asked to launch a ‘ball’ through three target rings by selecting the correct initial velocity. Observers chose one of 32 launch points on a computer screen that sends a dot along a quadratic ‘gravity-influenced’ trajectory. Each launch point corresponded to a different initial direction and velocity, visibly revealed by the angle and distance relative to the dot’s start point. One group received repetitive training on four ring configurations (4R) and a second receives variable training on 20 (20R); both completed the same number of total trials. Both groups underwent no-feedback pretest, posttest, and interleaved test sessions with new configurations to assess the effect of training experience.
While both groups performed similarly at pretest, performance changes at posttest for the two groups showed the predicted consequences of training. 4R observers learned to rely on the four launch points used during training, consistent with a learning strategy that fails to generalize to new stimuli. In contrast, 20R observers expanded their selection range in a manner consistent with an improved understanding of the overall relationship between launch point and trajectory shape.
Meeting abstract presented at VSS 2013